METHOD AND DEVICE FOR TESTING A TECHNICAL SYSTEM

A method for testing a technical system. The method includes: tests are carried out with the aid of a simulation of the system, the tests are evaluated with respect to a fulfillment measure of a quantitative requirement on the system and an error measure of the simulation, on the basis of the fulfillment measure and error measure, a classification of the tests as either reliable or unreliable is carried out.

Skip to: Description  ·  Claims  · Patent History  ·  Patent History
Description
CROSS REFERENCE

The present application claims the benefit under 35 U.S.C. § 119 of German Patent Application No. 102020205539.4 filed Apr. 30, 2020, which is expressly incorporated herein by reference in its entirety.

FIELD

The present invention relates to a method for testing a technical system. The present invention additionally relates to a corresponding device, a corresponding computer program, and a corresponding storage medium.

BACKGROUND INFORMATION

In software technology, using models to automate test activities and generate test artifacts in the test process is summarized under the generic term “model-based testing” (MBT). For example, generating test cases from models which describe the setpoint behavior of the system to be tested is conventional.

Embedded systems in particular are dependent on coherent input signals from sensors and in turn stimulate their surroundings by output signals to various actuators. In the course of the verification and preceding development phases of such a system, in a control loop its model (model in the loop, MiL), software (software in the loop, SiL), processor (processor in the loop, PiL), or overall hardware (hardware in the loop, HiL) is therefore simulated together with a model of the surroundings. In automotive engineering, simulators corresponding to this principle for testing electronic control units are sometimes referred to as component test stands, module test stands, or integration test stands depending on the test phase and test object.

German Patent Application No. DE 10303489 A1 describes such a method for testing software of a control unit of a vehicle, a power tool, or a robotics system, in which a control loop controllable by the control unit is at least partially simulated by a test system in that output signals are generated by the control unit and these output signals of the control unit are transferred to first hardware components via a first connection and signals from second hardware components are transferred as input signals to the control unit via a second connection, the output signals being provided as first control values in the software and additionally being transferred to the test system via a communication interface in real time with respect to the control path.

Such simulations are widespread in various areas of technology and are used, for example to test embedded systems for suitability in power tools, engine control units for drive systems, steering systems, and braking systems, camera systems, systems having components of artificial intelligence and machine learning, robotics systems, or autonomous vehicles in early phases of their development. Nonetheless, the results of simulation models according to the related art are only incorporated to a limited extent in release decisions because of a lack of trust in their reliability.

SUMMARY

The present invention provides a method for testing a technical system, a corresponding device, a corresponding computer program, and a corresponding storage medium.

In accordance with an example embodiment of the present invention, the quality of simulation models is decisive for the correct predictability of the test results thus achievable. In the field of MBT, the subdiscipline of validation relates to the object of comparing real measurements to simulation results. Various metrics, measures, or other comparators are used for this purpose, which link signals to one another and are to be referred to in summary hereinafter as signal metrics (SM). Examples of such signal metrics are metrics which compare size, phase shift, and correlations. Some signal metrics are defined by relevant norms, for example according to ISO 18571.

Expressed in more general terms, uncertainty quantification technologies assist the estimation of the simulation quality and model quality. The result of an assessment of the model quality using a signal metrics or more generally using an uncertainty quantification method for a certain input X, which may be a parameter or a scenario is denoted hereinafter as the simulation model error metric—in short: error metric—SMerrorX. For the generalization (interpolation and extrapolation) of SMerrorX for previously unobserved inputs, parameters, or scenarios X machine learning methods may be used, for example based on so-called Gaussian processes.

During the verification, the test subject (system under test, SUT) is typically studied on the basis of a requirement, specification, or performance indicator. It is to be noted that Boolean requirements or specifications may often be converted into quantitative measurements in that formalisms such as signal temporal logic (STL) are used. Such formalisms may be used as the basis of quantitative semantics, which represent a generalization of the verification insofar as a positive value indicates the fulfillment and a negative value indicates the infringement of a requirement. Such requirements, specifications, or performance measures are referred to in summary hereinafter as “quantitative requirements” (QSpec).

These types of quantitative requirements may be checked either on the basis of the real SUT or a model thereof—equivalent to a “virtual SUT.” For the purpose of this verification, catalogs are compiled having test cases which a SUT has to meet in order to decide whether it has the desired performance and safety properties. Such a test case may be parameterized and thus cover an arbitrary number of individual tests.

In accordance with an example embodiment of the present invention, the need for reliable test results is taken into consideration in order to ensure the performance and safety properties of a SUT. In particular when carrying out tests on the basis of a simulation of the system or a subcomponent—instead of the real system—it is to be ensured that the simulation results are trustworthy.

One goal of this approach is therefore to obtain such reliable test results on the basis of simulations that they may be used as a replacement for real test cases. The costs for the testing are thus to be reduced by reducing the number of actual experiments.

An array of tests, for example, a test catalog or a parametric test, which the SUT is to meet, are given here. An example embodiment of the present approach provides dividing the quantity of tests into two test sets: on the one hand, tests which have to be carried out on the real system and, on the other hand, tests which may be carried out on the basis of a simulation.

The provided approach enables the end user to be advised whether a virtual test is trustworthy or not. It enables a user to be advised when a test is to be carried out on the real system. It enables automatic initiation of the execution of a real test if the virtual test is not to be classified as reliable.

One advantage of the approach according to the present invention for this object is that in contrast to concepts which are exclusively based on validation or exclusively on verification, both approaches are unified in a skilled way. For this purpose, a “virtual test classifier” is introduced, which combines the requirements of model validation and product test. This is achieved by linking pieces of information from the validation of simulation and model quality (SMerrorX), on the one hand, and test requirements (QSpec), on the other hand.

Greatly varying fields come into consideration for the application of corresponding tests. For example, the functional reliability of automated systems is to be considered, as are used, for example, for automating driving functions (automated driving).

Advantageous refinements and improvements of the present invention are possible by way of the measures disclosed herein.

An automated, computer-implemented test environment may thus be provided in order to improve the quality of the tested hardware or software products largely automatically.

BRIEF DESCRIPTION OF THE DRAWINGS

Exemplary embodiments of the present invention are illustrated in the figures and explained in greater detail in the description below.

FIG. 1 shows a virtual test classifier, in accordance with an example embodiment of the present invention.

FIG. 2 shows a first approach for generating the decision limit of the classifier on the basis of data, in accordance with an example embodiment of the present invention.

FIG. 3 shows a second approach for generating the decision limit of the classifier on the basis of a formal solution, in accordance with an example embodiment of the present invention.

FIG. 4 shows the description of a method according to the present invention from an application viewpoint, in accordance with an example embodiment of the present invention.

FIG. 5 shows the visualization of a classification result in a feature space spanned by the test parameters, in accordance with an example embodiment of the present invention.

FIG. 6 schematically shows a control unit according to a second specific embodiment in accordance with the present invention.

DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS

According to the present invention, in the context of a test X, which may be taken as a test case from a test catalog or may be obtained as an instance of a parametric test, simulation model error SMerrorX is evaluated and quantitative specification QSpec is assessed on the basis of a simulation of the SUT. The virtual test classifier uses SMerrorX and QSpec as the input and makes a binary decision as to whether the test result based on the simulation is trustworthy or not.

According to the linguistic usage typical in information technology and in particular pattern recognition, a classifier is to be understood here as any algorithm or any mathematical function which maps a feature space on a set of classes which were formed and bounded from one another in the course of a classification. To be able to decide in which class an object is to be categorized or classed (colloquially also “classified”), the classifier uses so-called class or decision limits. If it is not important to distinguish between method and instance, the term “classifier” is used in technical language and also sometimes in the following as equivalent with “classification” or “classing.”

FIG. 1 illustrates such a classification in the present exemplary application. In this case, each point corresponds to a test which was carried out in the course of the simulation and for which fulfillment measure 13 of the requirement QSpec and error measure 14 SMerrorX were calculated. QSpec is defined in this case so that it assumes a positive value if it may be inferred from the tests that the system meets the respective requirement (reference numeral 24), and negative if the system fails the requirement (reference numeral 25).

As may be seen from the figure, decision limit 19 of classifier 18 divides the space into four classes A, B, C and D. Tests of class A were passed by the system with high reliability. For tests of classes B and C, the simulation only supplies unreliable results; such tests are therefore to be carried out on the real system. Tests of class D were failed on the system with high reliability.

This virtual test classifier 18 is based on the consideration that a requirement which is only barely met in the simulation may only replace the testing of the real system if at most a marginal model error 14 is to be presumed. On the other hand, in the event of a high absolute value of fulfillment measure 13 of quantitative requirement QSpec, i.e., in the case of a specification which is greatly exceeded or clearly failed, a certain deviation of the simulation results from corresponding experimental measurements may be accepted.

Since this way of viewing things presumes the knowledge of model error SMerrorX of the simulation model, it is presumed that the latter was subjected to a verification and validation before the use of virtual test classifier 18. Within the scope of this validation—for example on the basis of a Gaussian process or in another way by machine learning—a generalized model is to be formed which supplies SMerrorX for a given X. It is to be noted that the trustworthiness of the simulation is decisively dependent on the correctness of this generalized model.

FIG. 2 illustrates one possible approach for drawing decision limit 19 (FIG. 1) of classifier 18 on the basis of data. In the simplest case, limit 19 extends here along a line through the origin. The slope of the straight line is preferably to be selected so that all points in which fulfillment measure 13 of quantitative requirement QSpec changes the sign between simulation 11 and real measurement 21—thus quasi all tests 12 in which the simulation model fails—are in areas C and B and these areas are moreover as small as possible.

Furthermore, a more general, for example polynomial decision limit 19 comes into consideration, whose function curve is adapted with the aid of linear programming in such a way that it meets the criterion of a classifier 18 VTC. All points in which fulfillment measure 13 of quantitative requirement QSpec differs in sign between simulation 11 and real measurement 21—thus equivalent to all tests 12 in which the simulation model fails—are also in areas C and B in this case.

FIG. 3 illustrates the alternative approach of defining classifier 18 by solving 23 a formal equation system, on which the definition equations of fulfillment measure 13 and error measure 14 are based. The resulting function, which assigns a truth value to feature vector 13, 14 formed from these two measures, may alternately be specified deterministically or stochastically.

For the purposes of the following statements, I is the input quantity, O is the output quantity—under certain circumstances also including inputs, and m1,m2:I→O is the system model and real system as functions, which may only be observed for a finite number of inputs by simulation 11 or experimental measurement 21. Furthermore q:O×O→ is simulation model error SMerrorX, i.e., distance or error measure 14 of two outputs corresponding to one another. Finally Iϵ:={i|q(m1(i),m2(i))=ϵ} is the set of all inputs for which error measure (14) assumes value ϵ.

Starting from these definitions, the deviation of fulfillment measure 13 of a requirement for each input i ∈ Iϵ may be restricted on the upper end as follows by a term which is dependent neither on m1 nor on m2:

ϵ i I ϵ : p ( m 1 ( i ) ) - p ( m 2 ( i ) ) sup j I ϵ p ( m 1 ( j ) ) - p ( m 2 ( j ) ) sup ( o 1 , o 2 ) q - 1 ( ϵ ) p ( o 1 ) - p ( o 2 ) Formula 1

Classifier 18 therefore results as

VTC ( ϵ , δ ) = { W if δ > sup ( o 1 , o 2 ) q - 1 ( ϵ ) p ( o 1 ) - p ( o 2 ) F otherwise . Formula 2

The simulation model is classified as reliable here in the case of VTC(ϵ,δ)=W in the meaning that m1 and m2 correspond with respect to p. It is to be noted that classifier 18 requires the reversal of q.

One main advantage of this representation is that virtual test classifier 18 may be formulated independently of m1 and m2 since it is only dependent on fulfillment measure 13 of the quantitative requirement and error measure 14. Proceeding from a single error measure 14 and a plurality n of quantitative requirements, n virtual test classifiers 18 may thus be calculated, namely one for each requirement. The model is therefore only to be validated once with respect to error measure 14 and not for example with regard to each individual requirement.

This observation may be generalized in a simple way for a plurality m of error measures and a plurality n of quantitative requirements, m typically resulting as very small and und n resulting as large. In this case, n·m virtual test classifiers 18 may be calculated. If one of these classifiers 18 supplies value W, the simulation result may be considered to be reliable. This enables a more precise classification, since some error measures 14 may be more suitable for certain requirements than others.

Alternatively, virtual test classifier 18 may be defined in a stochastic framework, in which the inputs are assumed to be randomly distributed—according to an arbitrary probability density function. For this purpose, Fϵ(δ):=P(|p(m1(i))−p(m2(i))|≤δ|q(m1(i),m2(i))=ϵ) denotes the conditional cumulative distribution function of the deviation of fulfillment measure 13 under the assumption that error measure 14 assumes value ϵ. At a threshold value τ ∈ (0,1) of the probability that classifier 18 makes the correct decision—value τ is therefore typically close to 1—, virtual test classifier 18 may be defined as follows:

VTC ( ϵ , δ ) = { W if δ > inf F ϵ - 1 ( τ ) F otherwise Formula 3

FIG. 4 shows a method 10 according to the present invention from an application viewpoint under the following assumptions:

    • A simulation model 11 and a set of tests 12 together with defined inputs are predetermined.
    • Requirements QSpec are quantifiable and predefined and are implemented within the scope of a monitoring system which evaluates tests 12 with respect to fulfillment measure 13 of these requirements. In the figure, both fulfillment measures 13 relate to the same requirement QSpec, but evaluated once on the basis of simulation 11 and once in the course of experimental measurement 21 on the system.
    • SMerrorX is an error measure 14 which was defined beforehand. For some test inputs, simulation 11 and measurement 21 were thus already carried out, and error measure 14 generalizes corresponding tests 12 to new experiments which have not yet been carried out with a certain reliability, which is determined, for example, by an upper and lower limit for error measure 14. For classifier 18 (FIGS. 1 through 3), only the most unfavorable, i.e., the highest error measure 14, is used. It is to be noted that classifier 18 may be used to further refine error measure 14.

Under these assumptions, method 10 may be designed as follows:

    • 1. Classifier 18 is defined according to the above explanations.
    • 2. Tests 12 are carried out with the aid of simulation 11 and experimental measurement 21, output signals being generated.
    • 3. The output signals are evaluated with respect to fulfillment measure 13 of requirements QSpec and error measure 14 of simulation 11 according to SMerrorX error model.
    • 4. Fulfillment measure 13 and error measure 14, which are taken, on the one hand, in simulation 11 and, on the other hand, in measurement 21, are supplied to classifier 18.
    • 5. For each test 12, classifier 18 carries out a classification 15 into one of the following classes A, B, C, D (FIG. 1): Test 12 was successful in simulation 11 and its result is reliable 16; the test failed in simulation 11 and its result is reliable 16; or the result of simulation 11 is unreliable 17.
    • 6. Reliable 16 test results, for which simulation 11 is now considered to be trustworthy, are added to a corresponding database 31.
    • 7. Unreliable 17 tests 12 may be taken as a reason to recommend 32 to the user that a corresponding measurement 21 be carried out on the system.
    • 8. Such a measurement (21) may optionally be initiated manually or automatically.
    • 9. If decision limit 19 (FIG. 1) of classifier 18 was drawn on the basis of data as explained with respect to FIG. 2, it may—also optionally—be updated on the basis of the results of measurement 21 and improved 33 with the aid of these data points.

FIG. 5 sketches the possible visualization of a classification result in a feature space of the test parameters spanned by the test parameters (in the following: “parameter space”). For certain parameters 26, 27 of a test 12—exemplary according to the figure, distance 26 and mass 27 of a vehicle merging into the ego lane—fulfillment measure 13 and error measure 14 are each represented as points in the parameter space. In a virtual test environment 29, visualization 28 of classification 15 of tests 12 is carried out by classifier 18 in the parameter space.

This example method 10 may be implemented, for example, in software or hardware or in a mixed form of hardware and software, for example in a workstation 30, as the schematic view of FIG. 6 illustrates.

Claims

1. A method for testing a technical system, comprising the following steps:

carrying out tests using a simulation of the system;
evaluating the tests with respect to a fulfillment measure of a quantitative requirement on the system and an error measure of the simulation; and
based on the fulfillment measure and the error measure, carrying out a classification of the tests as either reliable or unreliable.

2. The method as recited in claim 1, wherein the technical system is an at least semi-autonomous robot or vehicle.

3. The method as recited in claim 1, wherein:

the classification is carried out by a classifier on based on a feature vector; and
the fulfillment measure and the error measure form components of the feature vector.

4. The method as recited in claim 3, wherein:

the classifier maps the feature vector on one of multiple classes; and
the classification takes place within predefined decision limits between the classes.

5. The method as recited in claim 4, wherein:

in a preparation phase, the simulation is confirmed by experimental measurement on the system; and
the decision limits are drawn in such a way that the fulfillment measure taken, on the one hand, in the simulation and, on the other hand, in the measurement deviates as little as possible.

6. The method as recited in claim 5, wherein further tests to be carried out in the preparation phase are selected automatically.

7. The method as recited in claim 4, wherein:

the classifier is defined by solving an equation system; and
the equation system includes definition equations of the fulfillment measure and the error measure.

8. The method as recited in claim 1, wherein the evaluation is carried out in such a way that the fulfillment measure is positive when the system meets the requirement, and negative when the system fails the requirement.

9. The method as recited in claim 1, wherein:

for certain parameters of each test, the fulfillment measure and the error measure are each represented in a feature space spanned by the parameters; and
after the evaluation, the classification is visualized in the feature space.

10. The method as recited in claim 1, wherein an automatic improvement of errors of the system recognized by the testing takes place.

11. A non-transitory machine-readable storage medium on which is stored a computer program for testing a technical system, the computer program, when executed by a computer, causing the computer to perform the following steps:

carrying out tests using a simulation of the system;
evaluating the tests with respect to a fulfillment measure of a quantitative requirement on the system and an error measure of the simulation; and
based on the fulfillment measure and the error measure, carrying out a classification of the tests as either reliable or unreliable.

12. A device configured to test a technical system, the device configured to:

carry out tests using a simulation of the system;
evaluate the tests with respect to a fulfillment measure of a quantitative requirement on the system and an error measure of the simulation; and
based on the fulfillment measure and the error measure, carry out a classification of the tests as either reliable or unreliable.
Patent History
Publication number: 20210342239
Type: Application
Filed: Mar 10, 2021
Publication Date: Nov 4, 2021
Patent Grant number: 11243858
Inventors: Christoph Gladisch (Renningen), Daniel Seiler-Thull (Stuttgart), Joachim Sohns (Ludwigsburg), Philipp Glaser (Stuttgart), Thomas Heinz (Stuttgart), Ji Su Yoon (Leonberg)
Application Number: 17/197,548
Classifications
International Classification: G06F 11/263 (20060101); G06F 11/22 (20060101); G06F 11/26 (20060101); G06F 11/07 (20060101); G06K 9/62 (20060101);